OpenAI's AI Found 948 Holes in the Linux Kernel and Wrote Patches for $1.2 Million

Picture an ordinary Linux server. A user with minimal privileges is logged in, but a bug in the kernel lets them become an administrator and take control of the entire machine. These bugs are called LPE, short for local privilege escalation.

OpenAI says its researchers built an autonomous pipeline that hunts for such holes in the Linux kernel, writes working exploits for them and prepares fixes. In a single run, a not-yet-released model confirmed 948 vulnerabilities and prepared patches for them, and the API bill came to about $1.2 million.

For comparison, an OpenAI employee says researchers typically find about 20 such bugs a year. If you run Linux servers, containers or cloud infrastructure, this is a signal that the pace of vulnerability discovery is changing.

What exactly OpenAI did

The pipeline covers the whole cycle: finding a vulnerability, exploiting it and patching it. It does not simply report "something may be wrong here." It checks that the hole can actually be exploited, and only then proposes a fix. That matters for kernel maintainers, who usually drown in false positives from automated scanners.

OpenAI also claims that its agents found and used at least 10 Linux kernel exploitation techniques that the team and invited experts consider new. According to the company, this shows the models are not just repeating known methods but inventing their own. This is OpenAI's own claim, and there is no independent verification yet.

How much it costs

The total API bill is about $1.2 million. Greg Brockman puts the figure at roughly $1 000 per vulnerability; simply dividing $1.2 million by 948 gives about $1 300. Eric Wallace says the model found about a thousand LPEs in a week, spending roughly $1 million on tokens. He adds that OpenAI is now directing a "significant amount" of frontier compute to fixing vulnerabilities in open-source code for free.

The trend is telling, too. In one comparison, GPT-5.6 Sol found 166 vulnerabilities in July 2026 for about $500 thousand, or around $3 000 each. The new model closed 948 holes for $1.2 million: almost six times as many bugs, with the cost per find falling by more than half. One commenter believes the real internal cost is even lower than the API list price, but that is an unconfirmed estimate.

What to do in practice

First, OpenAI is working with Linux maintainers and its partners at Patch the Planet on reports and patches. Expect a wave of kernel updates, and iron out your process for rolling them out quickly across your servers.

Second, if a model can find a thousand LPEs in a week, similar capabilities will sooner or later reach attackers as well. The window between a bug appearing and being exploited is shrinking, so fast updates, workload isolation and the principle of least privilege are moving from checklist items to necessities.

Third, for those building security tools: the bar has gone up. The value no longer lies in a scanner that shows suspicions, but in an agent that proves the vulnerability and delivers a ready patch.

Caveats

The model used has not been published and is separate from the rumors about GPT-6.2 Sol and Astra, so outsiders cannot reproduce the result. All figures come from OpenAI and its employees. It is unclear how many of the 948 findings maintainers will judge critical, and how many will turn out to be low-risk or duplicates. Even with those caveats, the takeaway matters: finding kernel vulnerabilities increasingly depends on compute budget, not just on the number of experts.